Planning device, planning method, and planning program

Abstract

This planning device comprises: an overall automated planning problem setting unit that generates, in a formal language, an automated planning problem from a task instruction, an environment graph, and domain information; a task breakdown / assignment setting unit that has a generative AI that divides an overall task into subtasks on the basis of the automated planning problem, and also assigns the subtasks and allocates a computational priority order to each agent; a partial automated planning problem setting / evaluation unit that has a generative AI that generates a partial automated planning problem of an agent for executing the assigned subtasks; an automated planning unit that solves the partial automated planning problem by means of an automated planning solver, and computes a state transition / action sequence of the agent; and a cost prediction unit that converts the state transition of the agent into a graph representation, and performs a cost prediction using a machine learning model for predicting the feasibility and cost of an action sequence of the agent on the basis of said information.

Description

Planning Device, Planning Method, and Planning Program

[0001] The present invention relates to a planning device, a planning method, and a planning program.

[0002] In recent years, technologies have been developed for automating tasks such as pick-and-place by agents that perform work on behalf of users. Therefore, development has also been underway for technologies to calculate the work plans of agents.

[0003] For example, Non-Patent Document 1 describes a method for calculating an operation plan for a single-agent system by utilizing Large Language Models (LLMs) and Planning Domain Definition Language (PDDL), which is one of the formal languages for automatic planning, based on task instructions in natural language and environmental information expressed by a 3D Scene Graph (3DSG). In addition, Non-Patent Document 2 describes a method for calculating code for controlling multi-agents from natural language instructions by utilizing LLMs for a multi-agent system.

[0004] Y. Liu, et al,: “DELTA: Decomposed Efficient Long-Term Robot Task Planning using Large Language Models”, Arxiv, 2024.S. S. Kannan, et al.: “SMART-LLM: Smart Multi-Agent Robot Task Planning using Large Language Models”, Arxiv, 2024.

[0005] Note that each disclosure of the above prior art documents is incorporated herein by reference. The following analysis is made by the inventors of the present invention.

[0006] Incidentally, artificial intelligence such as LLM (Limited Licensing Model) exhibits a phenomenon called hallucination, which generates information not based on facts. In conventional LLM-based agent task planning methods, hallucination can lead to inconsistent procedures and the calculation of impossible task plans. Furthermore, because the feasibility of each action at the control level is not considered during planning, task plans that are physically impossible to execute, even if the procedures are consistent, can be calculated. In addition, conventional LLM-based agent task planning methods have not adequately considered their application to large-scale systems consisting of multiple agents, resulting in the inability to plan efficient tasks that take into account agent cooperation and interference avoidance.

[0007] The object of the present invention is to provide a planning device, a planning method, and a planning program that contribute to calculating a plan that considers optimality while ensuring feasibility, in view of the above-mentioned problems.

[0008] In a first aspect of the present invention, a planning device is provided comprising: an overall automatic planning problem setting unit that generates an automatic planning problem in formal language from task instructions, an environment graph, and domain information; a task decomposition / assignment setting unit having a generation AI that divides the overall task into subtasks based on the automatic planning problem and assigns the subtasks and calculation priorities to each agent; a partial automatic planning problem setting / evaluation unit having a generation AI that generates a partial automatic planning problem for agents to execute the assigned subtasks; an automatic planning unit that solves the partial automatic planning problem using an automatic planning solver and calculates the state transition / action sequence of the agents; and a cost prediction unit that converts the state transitions of the agents into a graph representation and performs cost prediction using a machine learning model that predicts the feasibility and cost of the agent's action sequence based on that information.

[0009] A second aspect of the present invention provides a planning method comprising: an overall automatic planning problem setting process that generates an automatic planning problem in formal language from task instructions, an environment graph, and domain information; a task decomposition / assignment setting process having a generative AI that divides the overall task into subtasks based on the automatic planning problem and assigns the subtasks and calculation priorities to each agent; a partial automatic planning problem setting / evaluation process having a generative AI that generates a partial automatic planning problem for agents to execute the assigned subtasks; an automatic planning process that solves the partial automatic planning problem using an automatic planning solver and calculates the state transition / action sequence of the agents; and a cost prediction process that converts the state transitions of the agents into a graph representation and performs cost prediction using a machine learning model that predicts the feasibility and cost of the agent's action sequence based on that information.

[0010] A third aspect of the present invention provides a planning program that causes a computer to execute the following: an overall automatic planning problem setting process that generates an automatic planning problem in formal language from task instructions, an environment graph, and domain information; a task decomposition / assignment setting process having a generative AI that divides the overall task into subtasks based on the automatic planning problem and assigns the subtasks and calculation priorities to each agent; a partial automatic planning problem setting / evaluation process having a generative AI that generates a partial automatic planning problem for agents to execute the assigned subtasks; an automatic planning process that solves the partial automatic planning problem using an automatic planning solver and calculates the state transitions / action sequences of the agents; and a cost prediction process that converts the state transitions of the agents into a graph representation and performs cost prediction using a machine learning model that predicts the feasibility and cost of the action sequences of the agents based on that information. This program can be recorded on a computer-readable storage medium. The storage medium can be a non-transitory material such as semiconductor memory, hard disk, magnetic recording medium, or optical recording medium. The present invention can also be embodied as a computer program product.

[0011] From each perspective of the present invention, it is possible to provide a planning device, a planning method, and a planning program that contribute to calculating a plan that considers optimality while ensuring feasibility.

[0012] Figure 1 is a schematic diagram of the planning device according to one embodiment. Figure 2 is a diagram illustrating the processing of the overall automatic planning problem setting unit. Figure 3 is a diagram illustrating the processing of the task decomposition / assignment setting unit (facilitator LLM) and the partial automatic planning problem setting / evaluation unit group (agent LLM). Figure 4 is a diagram illustrating the processing of the partial automatic planning problem setting / evaluation unit group (agent LLM), the automatic planning unit, and the cost forecasting unit. Figure 5 is a flowchart showing the procedure of the planning method according to one embodiment. Figure 6 is a diagram showing an example of the configuration of the cost forecasting unit. Figure 7 is a diagram showing an example of the processing of the graph transition data generation unit. Figure 8 is a diagram showing an example of the processing of the graph transition cost forecasting unit. Figure 9 is a flowchart showing the procedure of the cost forecasting unit. Figure 10 is a diagram showing an example of a situation in which the planning device and planning method according to one embodiment can be used. Figure 11 is a diagram showing an example of the hardware configuration of the planning device used in the embodiment.

[0013] Embodiments of the present invention will be described below with reference to the drawings. However, the present invention is not limited to the embodiments described below. In each drawing, the same or corresponding elements are appropriately denoted by the same reference numerals. Furthermore, it should be noted that the drawings are schematic, and the dimensional relationships and ratios of each element may differ from those of reality. Even between drawings, there may be parts where the dimensional relationships and ratios differ from each other.

[0014] Figure 1 is a schematic diagram of a planning device according to one embodiment. As shown in Figure 1, the planning device 10 comprises an overall automatic planning problem setting unit 11, a task decomposition / assignment setting unit (facilitator LLM) 12, a group of partial automatic planning problem setting / evaluation units (agent LLMs) 131, ..., 13N, an automatic planning unit group 141, ..., 14N, a cost prediction unit group 151, ..., 15N, and a planning result database 16. Note that the planning device 10 shown in Figure 1 is assumed to create work plans for N agents.

[0015] The overall automated planning problem setting unit 11 generates an automated planning problem in formal language from task instructions, environment graphs, and domain information. Here, task instructions are work instructions to be performed by agents, environment graphs are graphical representations of the positions, objects, and agents that agents can move to, domain information is information indicating state transition rules for agents and objects, and the automated planning problem is a problem text for the automated planning solver that describes in formal language (PDDL, etc.) the problem of calculating the state transition / action sequence of all agents to execute a given task, and includes information on task instructions, environment graphs, and domain information in formal language format. The task decomposition / assignment setting unit (facilitator LLM) 12 has a generation AI and divides the overall task into subtasks based on the automated planning problem, and assigns the subtasks and calculation priorities to each agent. The partial automated planning problem setting / evaluation unit group (agent LLM) 131, . . . 13N has a generation AI and generates partial automated planning problems for agents to execute the assigned subtasks. A partial automated planning problem is a problem text for an automated planning solver that describes, in a formal language, the problem of calculating the state transition / action column for each agent to execute an assigned subtask, similar to an automated planning problem. It includes information about subtasks, environment graphs, and domain information in a formal language format. The automated planning units 141, . . . , 14N solve the partial automated planning problem using an automated planning solver and calculate the state transition / action column for the agents. The cost prediction units 151, . . . , 15N convert the agent state transitions into a graph representation and perform cost prediction using a machine learning model that predicts the feasibility and cost of the agent's action column based on that information. The planning result database 16 stores the agent planning results created by the automated planning units 141, . . . , 14N. The automated planning units 141, . . . , 14N save the agent's current state transition column, action column, and predicted cost column in the planning result database 16, and the partial automated planning problem setting / evaluation units (agent LLM) 131, . ..., output to 13N.

[0016] Here, as will be described later, the partially automated planning problem setting / evaluation units (agent LLMs) 131, ..., 13N evaluate the current plan based on the agent's current state transition column, action column, and predicted cost column, and output the evaluation result to the task decomposition / assignment setting unit (facilitator LLM) 12. The task decomposition / assignment setting unit (facilitator LLM) 12 then modifies the subtasks based on this evaluation result. Meanwhile, the partially automated planning problem setting / evaluation units (agent LLMs) 131, ..., 13N correct the partially automated planning problem according to the subtask modifications made by the task decomposition / assignment setting unit (facilitator LLM) 12. In this way, the planning device 10 improves the agent's work plan while interacting with the task decomposition / assignment setting unit (facilitator LLM) 12 and the partially automated planning problem setting / evaluation units (agent LLMs) 131, ..., 13N.

[0017] Figure 2 illustrates an example of the processing of the overall automated programming problem setting unit. As shown in Figure 2, the overall automated programming problem setting unit 11 generates an automated programming problem in formal language from task instructions, an environment graph, and domain information. Here, the environment graph can be represented using 3DSG, the domain information is, for example, information indicating state transition rules and can be represented using PDDL, and the formal language of the automated programming problem can also be represented using PDDL. Domain information may also be automatically generated from natural language information using LLM.

[0018] Figure 3 illustrates an example of the processing of the Task Decomposition / Assignment Setting Unit (Facilitator LLM) and the Partial Automated Planning Problem Setting / Evaluation Units (Agent LLM). As shown in Figure 3, the prompt input to the Task Decomposition / Assignment Setting Unit (Facilitator LLM) 12 incorporates instructions on the output method of the Task Decomposition / Assignment Setting Unit (Facilitator LLM) 12, as well as domain information and automated planning problem information. The Task Decomposition / Assignment Setting Unit (Facilitator LLM) 12, in accordance with the input prompt, divides the overall task into subtasks based on the automated planning problem and assigns subtasks and calculation priorities to each agent. Meanwhile, the Partial Automated Planning Problem Setting / Evaluation Units (Agent LLM) 131, ..., 13N generate partial automated planning problems for agents to execute the assigned subtasks. The partial automated planning problems are generated by the Automated Planning Units 141, ..., 14N and the Cost Prediction Units 151, ... , 15N is used to calculate the plan for each agent, and the plan results for each agent are evaluated by the Partial Automatic Planning Problem Setting / Evaluation Unit (Agent LLM) 131, , , 13N and fed back to the Task Decomposition / Assignment Setting Unit (Facilitator LLM) 12 as suggested improvements to each agent's plan. The suggested improvements to each agent's plan are reflected in subsequent prompts to improve the subtask assignment and calculation priority output by the Task Decomposition / Assignment Setting Unit (Facilitator LLM) 12.

[0019] Figure 4 illustrates an example of the processing of the Partial Automated Planning Problem Setting / Evaluation Unit (Agent LLM), the Automated Planning Unit, and the Cost Forecasting Unit. As shown in Figure 4, the prompt input to the Partial Automated Planning Problem Setting / Evaluation Unit (Agent LLM) 13i incorporates the assigned task, domain information, and information about the automated planning problem. The Partial Automated Planning Problem Setting / Evaluation Unit (Agent LLM) 13i generates a partial automated planning problem for the agent to execute the assigned subtask, according to the input prompt. The Automated Planning Unit 14i solves the partial automated planning problem using an automated planning solver and calculates the agent's state transition / action sequence. The automated planning solver can utilize a general-purpose solver such as Fast Downward. Furthermore, the Cost Forecasting Unit converts the agent's state transitions into a graph representation and performs cost forecasting using a machine learning model that predicts the feasibility and cost of the agent's action sequence based on that information. The agent's calculation results are reflected in the prompts input to the Partial Automated Design Problem Setting / Evaluation Unit (Agent LLM) 13i, improving the Partial Automated Design Problem Output by the Partial Automated Design Problem Setting / Evaluation Unit (Agent LLM) 13i.

[0020] Figure 5 is a flowchart showing the procedure of a planning method according to one embodiment. As shown in Figure 5, the planning method is assumed to be executed by an overall automatic planning problem setting unit, a task decomposition / assignment setting unit, a group of partial automatic planning problem setting / evaluation units, an automatic planning unit, and a cost forecasting unit, but the implementation of the planning method is not limited to a specific configuration.

[0021] First, the overall automated planning problem setting unit generates an automated planning problem in formal language from task instructions, environment graphs, and domain information (step S1). Next, the task decomposition / assignment setting unit divides the overall task into subtasks based on the automated planning problem and assigns each agent the subtask and its calculation priority (step S2). The partial automated planning problem setting / evaluation unit group generates a partial automated planning problem for the agents to execute the assigned subtasks (step S3). The automated planning unit solves the partial automated planning problem using an automated planning solver and calculates the agent's state transition / action sequence (step S4). The cost prediction unit converts the agent's state transitions into a graph representation and performs cost prediction using a machine learning model that predicts the feasibility and cost of the agent's action sequence based on that information (step S5). After step S5, the termination condition is determined, and the process from step S3 to step S5 is repeated until the termination condition is met.

[0022] Subsequently, the automated planning unit saves the agent's current state transition column, action column, and predicted cost column to the planning results database and outputs them to the partial automated planning problem setting / evaluation unit group (step S6). The partial automated planning problem setting / evaluation unit group then evaluates the current plan based on the agent's current state transition column, action column, and predicted cost column, and outputs the evaluation results to the task decomposition / assignment setting unit (step S7). The task decomposition / assignment setting unit modifies the subtasks based on these evaluation results (step S2). Meanwhile, the partial automated planning problem setting / evaluation unit group corrects the partial automated planning problem according to the subtask modifications made by the task decomposition / assignment setting unit 12 (step S3). In this way, the planning method improves the agent's work plan through dialogue between the task decomposition / assignment setting unit and the partial automated planning problem setting / evaluation unit group.

[0023] Figure 6 shows an example of the configuration of the cost prediction unit. As shown in Figure 6, the cost prediction unit 15i includes a graph transition data generation unit 151i, a graph transition cost prediction unit 152i, and a graph transition cost uncertainty evaluation unit 153i. The graph transition data generation unit 151i represents the state change due to an agent's action as a graph (state transition before and after graph) based on the state and action of the agent calculated by the automatic planning unit and the calculated state and action information of other agents. The graph transition cost prediction unit 152i takes the obtained state transition before and after graph as input and predicts the cost related to the agent's action. The graph transition cost uncertainty evaluation unit 153i takes the obtained cost prediction value as input and outputs a cost value that reflects the uncertainty of the prediction.

[0024] Figure 7 shows an example of the processing performed by the graph transition data generation unit. As shown in Figure 7, the graph transition data generation unit converts the state before and after an action, expressed in a formal language, into a graph representation. For example, when an Agent performs an action to move from a location Place 1 to a location Place 2, as shown in Figure 7, the information on the state before and after the action, expressed in a formal language, is converted into a graph representation of the state before and after the action.

[0025] Figure 8 shows an example of the processing of the graph transition cost prediction unit. As shown in Figure 8, the graph transition cost prediction unit takes the graphs before and after state transitions and action data as input to predict the cost of an agent's action. The graph transition cost prediction unit can use a machine learning model that has been trained in advance. The example shown in Figure 8 is an example of processing using a machine learning model using Graph Neural Network. The graph transition cost prediction unit also determines the feasibility of the action and outputs an impossible action with a cost of ∞ or -1, etc.

[0026] The graph transition cost uncertainty evaluation unit utilizes methods such as an ensemble of predicted values ​​and Conformal Prediction to output a cost value that reflects the prediction uncertainty of the graph transition cost prediction unit. For example, when evaluating prediction uncertainty using an ensemble of predicted values, the graph transition cost prediction unit performs multiple cost predictions and calculates the mean μ and standard deviation σ of the obtained predicted costs. From these values, the unit outputs a cost prediction value that reflects the prediction uncertainty by adding the product of the non-negative scalar value confidence interval parameter β and the standard deviation σ, and the mean μ.

[0027]

[0028] Figure 9 is a flowchart showing the processing steps of the cost forecasting unit. As shown in Figure 9, the processing of the cost forecasting unit is assumed to be performed by a graph transition data generation unit, a graph transition cost forecasting unit, and a graph transition cost uncertainty evaluation unit, but the processing of the cost forecasting unit is not limited to a specific configuration.

[0029] First, the graph transition data generation unit generates graph transition data from the formal language state (step S51). Next, the graph transition cost prediction unit inputs the graph generation data into a machine learning model and predicts the graph transition cost (step S52). Then, the graph transition cost uncertainty evaluation unit evaluates the uncertainty of the graph transition cost prediction value and outputs the corrected cost prediction value (step S53).

[0030] Figure 10 shows an example of a situation in which a planning device and planning method according to one embodiment can be used. The example situation shown in Figure 10 is a picking operation in which three robot agents pick up target objects from shelves and store them on shelves for shipping by a specified time. Each agent can perform three types of actions: "move," "pick," and "place," and each action has the condition that the time required to perform it differs depending on the situation. For example, the longer the distance to move, the longer the time required, and the picking time differs depending on the object. Furthermore, there are cases in which an action cannot be performed depending on the situation. For example, the aisle between shelves is too narrow for the agents to pass each other.

[0031] Specifically, the example shown in Figure 10 is a scenario for a task instruction where, in an environment graph consisting of movable positions, shelves, objects, and agents, all objects to be picked are stored on the shipping shelves by the final time (within 10 minutes), under the conditions of domain information that defines the agent's state (location), the object's state (location, which agent holds it, or which shelf it is stored on), and the actions all agents can take (move, pick, place) and the results thereof.

[0032] In this example scenario, the overall automated planning problem is a PDDL defining the picking work planning problem for all agents, while the partial automated planning problem is a PDDL defining the picking work planning problem for agent i, based on the subtask assigned to agent i. In this example scenario, the situation where some of the objects to be picked are stored in the shipping shelves corresponds to a subtask, and the predicted cost corresponds to the execution time of each action of the agent. Furthermore, the agent's plan evaluation includes results such as whether the calculated action is executable, whether the time constraints are met, and suggestions for improving task assignment.

[0033] The planning apparatus and planning method according to the embodiment can calculate a work plan for picking a required object, taking into account the coordination and interference avoidance of multiple agents, and aiming to complete the task in the shortest possible time while meeting time constraints.

[0034] (Example Hardware Configuration) Figure 11 shows an example hardware configuration of a planning device used in the embodiment. That is, the planning device 10 can realize each function of the planning device 10 by having the planning method described above executed as a program on an information processing device (computer) 20 that employs the hardware configuration shown in Figure 11. However, the example hardware configuration shown in Figure 11 is just one example of a hardware configuration that realizes each function of the planning device 10, and is not intended to limit the hardware configuration of the planning device 10. The planning device 10 may include hardware not shown in Figure 11.

[0035] As shown in Figure 11, the hardware configuration that the planning device 10 may employ includes, for example, a CPU (Central Processing Unit) 21, a main memory 22, an auxiliary memory 23, and an IF (Interface) unit 24, all interconnected by an internal bus.

[0036] The CPU 21 executes each instruction included in the planning program run by the information processing device (computer) 20. The main memory 22 is, for example, RAM (Random Access Memory) and temporarily stores various programs, such as the planning program run by the information processing device (computer) 20, for processing by the CPU 21.

[0037] The auxiliary storage device 23 is, for example, an HDD (Hard Disk Drive) and is capable of storing various programs, such as planning programs executed by the information processing device (computer) 20, for medium to long term. These various programs, such as planning programs, can be provided as program products recorded on a non-transitory computer-readable storage medium.

[0038] The IF unit 24 provides, for example, an interface for the input and output of the planning device 10.

[0039] The information processing device (computer) 20, which employs the hardware configuration described above, realizes each function of the planning device 10 by executing the aforementioned planning method as a program.

[0040] Furthermore, each disclosure of the above-mentioned non-patent literature cited shall be incorporated into this document by reference. Within the framework of the full disclosure of the present invention (including the claims), further modifications and adjustments to the embodiments or examples are possible based on the fundamental technical concept. Also, within the framework of the full disclosure of the present invention, various combinations or selections (including partial deletions) of various disclosure elements (including each element of each claim, each element of each embodiment or example, each element of each drawing, etc.) are possible. In other words, the present invention naturally includes the full disclosure, including the claims, and various modifications and alterations that a person skilled in the art could make in accordance with the technical concept. In particular, with respect to the numerical ranges described in this document, any numerical value or sub-range included within that range should be interpreted as being specifically described, even if not otherwise stated. Furthermore, each disclosure item of the above-mentioned cited documents may, if necessary, be used in part or in whole as part of the disclosure of the present invention, in accordance with the spirit of the present invention, and this is also considered to be included in the disclosure of this application.

[0041] 10 Planning device 11 Overall automatic planning problem setting unit 12 Task decomposition / assignment setting unit (facilitator LLM) 131, ..., 13N, 13i Partial automatic planning problem setting / evaluation unit (agent LLM) 141, ..., 14N, 14i Automatic planning unit 151, ..., 15N, 15i Cost prediction unit 151i Graph transition data generation unit 152i Graph transition cost prediction unit 153i Graph transition cost uncertainty evaluation unit 20 Information processing device 21 CPU 22 Main memory 23 Auxiliary memory 24 IF unit

Claims

1. A planning device comprising: an overall automatic planning problem setting unit that generates an automatic planning problem in formal language from task instructions, an environment graph, and domain information; a task decomposition / assignment setting unit having a generation AI that divides the overall task into subtasks based on the automatic planning problem and assigns the subtasks and calculation priorities to each agent; a partial automatic planning problem setting / evaluation unit having a generation AI that generates a partial automatic planning problem for agents to execute the assigned subtasks; an automatic planning unit that solves the partial automatic planning problem using an automatic planning solver and calculates the state transition / action sequence of the agents; and a cost prediction unit that converts the state transitions of the agents into a graph representation and performs cost prediction using a machine learning model that predicts the feasibility and cost of the agent's action sequence based on that information.

2. The planning device according to claim 1, wherein the automatic planning unit stores the current state transition column, action column, and predicted cost column of the agent in the planning results database and outputs them to the partial automatic planning problem setting / evaluation unit.

3. The planning device according to claim 2, wherein the partially automated planning problem setting / evaluation unit evaluates the current plan based on the current state transition column, action column, and predicted cost column of the agent, outputs the evaluation result to the task decomposition / assignment setting unit, and the task decomposition / assignment setting unit modifies the subtask based on the evaluation result.

4. The planning device according to claim 3, wherein the partial automatic planning problem setting / evaluation unit corrects the partial automatic planning problem in accordance with the modification of the subtasks performed by the task decomposition / assignment setting unit.

5. The planning apparatus according to claim 1, wherein the cost prediction unit comprises: a graph transition data generation unit that represents the state change due to an agent's action as a state transition before and after graph based on the state and action of an agent calculated by the automatic planning unit and information on the state and action of other agents that has been calculated; a graph transition cost prediction unit that predicts the cost related to the agent's action using the state transition before and after graph as input; and a graph transition cost uncertainty evaluation unit that outputs a cost value that reflects the uncertainty of the prediction using the predicted cost as input.

6. The planning device according to claim 5, wherein the graph transition cost prediction unit inputs the graphs before and after the state transition into a machine learning model using Graph Neural Network to predict the cost.

7. A planning method comprising: an overall automated planning problem setting process that generates an automated planning problem in formal language from task instructions, an environment graph, and domain information; a task decomposition / assignment setting process having a generative AI that divides the overall task into subtasks based on the automated planning problem and assigns the subtasks and calculation priorities to each agent; a partial automated planning problem setting / evaluation process having a generative AI that generates a partial automated planning problem for agents to execute the assigned subtasks; an automated planning process that solves the partial automated planning problem using an automated planning solver and calculates the state transitions / action columns of the agents; and a cost prediction process that converts the state transitions of the agents into a graph representation and performs cost prediction using a machine learning model that predicts the feasibility and cost of the action columns of the agents based on that information.

8. A planning program that causes a computer to execute: an overall automated planning problem setting process that generates an automated planning problem in formal language from task instructions, environment graphs, and domain information; a task decomposition / assignment setting process having a generative AI that divides the overall task into subtasks based on the automated planning problem and assigns the subtasks and calculation priorities to each agent; a partial automated planning problem setting / evaluation process having a generative AI that generates a partial automated planning problem for agents to execute the assigned subtasks; an automated planning process that solves the partial automated planning problem using an automated planning solver and calculates the state transitions / actions of the agents; and a cost prediction process that converts the state transitions of the agents into a graph representation and performs cost prediction using a machine learning model that predicts the feasibility and cost of the agent's action sequences based on that information.

Images